Speaker Names Attribution, Ettin 150M

An encoder-only span model that extracts the printed real name of each diarized speaker (Speaker A, Speaker B, and similar) from a transcript. It abstains when the transcript does not identify the speaker.

The model is based on jhu-clsp/ettin-encoder-150m with a SQuAD-style start/end span head. Position zero is the null answer.

Results

Evaluation uses a manually checked set of 264 transcript-speaker rows, including 131 named and 133 unnamed speakers.

Person F1 accepts any printed form of the correct person's name, for example Kevin for Kevin Cork. Exact-text F1 requires the full target text.

On the automatically verified held-out test split (638 rows), person F1 is 0.882 and exact-text F1 is 0.830.

Usage

This repository includes the small custom span head and inference code:

pip install "torch>=2.4" "transformers>=4.51,<5" huggingface_hub
huggingface-cli download hmarchant/speaker-names-attr-ettin-150m \
  --local-dir speaker-names-attr

python -m speaker_names.infer transcript.txt \
  --model speaker-names-attr

Input must be a diarized transcript containing labels such as:

[Speaker A]: I'm Sarah Jones. Tom Baker joins us now. Finance Minister David Cole announced the new budget earlier. Good morning, Tom.
[Speaker B]: Good morning, Sarah. David's announcement was unexpected.
[Speaker C]: Here is the weather forecast.

The command prints a JSON mapping from each speaker label to the extracted name or null.

Example output:

{
  "Speaker A": "Sarah Jones",
  "Speaker B": "Tom Baker",
  "Speaker C": null
}

Training

• 149M parameters
• 4,988 train, 550 validation, and 638 test rows
• maximum sequence length 4,096, stride 1,024
• batch size 1 with 16-step gradient accumulation
• SDPA attention on unpadded windows
• selected at epoch 3 by person-level validation F1
• decision threshold: -2.25

Labels came from grok-4.7, were grounded against verbatim transcript spans, then checked by a second model pass. Organisation names, programme names, and sentence fragments were removed.

Limitations

• The model only returns names printed in the transcript. It does not infer identities from voice, outside knowledge, or unprinted context.
• Speaker labels are assumed to be local to each transcript window.
• Accuracy depends on diarization and ASR quality.
• The training material is mostly English-language news and current-affairs transcripts, so performance may be lower in other domains and languages.
• The custom checkpoint is speaker_span.pt; it is not a standard AutoModelForQuestionAnswering checkpoint.